This invention relates to memory systems comprised of hub devices connected to a memory controller by a daisy chained channel. The hub devices are attached to, or reside upon, memory modules that contain memory devices. More particularly, this invention relates to the flow control of read data and the identification of read data returned to the controller by each hub device.
Many high performance computing main memory systems use multiple fully buffered memory modules connected to a memory controller by one or more channels. The memory modules contain a hub device and multiple memory devices. The hub device fully buffers command, address and data signals between the memory controller and the memory devices. The flow of read data is controlled using either a leveled latency or position dependant latency technique. In both cases, the memory controller is able to predict the return time of read data requested from the memory modules and to schedule commands to avoid collisions as read data is merged onto the controller interface by each memory module.
In some cases, the memory controller is able to issue a read data delay adder along with the read command. This instructs the targeted hub device to add additional delay to the return of read data in order to simplify the issuing of commands and to avoid collisions. In all cases, the read data must be returned in the order in which it was requested. Further, the total read data latency must be completely predictable by the memory controller. During run time operations, these two restrictions result in additional gaps being added to packets of read data that are returned from the memory modules. This adds latency to the average read operation. In addition, hubs are not able to use indeterminate techniques to return read data faster or slower than normal. These techniques include, but are not limited to, caching read data locally, reading memory devices speculatively, independently managing memory device address pages, data compression, etc.
To optimize average read data latency under real workload conditions, and to enable advanced hub device capabilities, what is needed is a way to allow memory modules to return read data to the memory controller at an unpredicted time. This must be done in a way that does not corrupt read data and that allows the memory controller to identify each read data packet. Preventing data corruption by avoiding data collisions is especially complicated as hub devices merge local read data onto a cascaded memory controller channel.